icegame

intro

Designed as an environments interacting with spin ice system.

Compile & Install

compile

sh compile.sh

will generate icegame.so in build/src. please move icegame.so to where you executing experiment

mv build/src/icegame.so /where/you/run/icegame

gym-icegame

Install python interface

python setup.py install

which depends on openai gym.

Game Scenario

Draw a proposed loop, then summit.

actions

• Directional action: (up, down, left, right, next up, next down) 6 operations in total

Two options:

• Metropolis button:
• Loop Auto-detection:

function called by env.step(action_index)

• step()

rewards

• Step-wise reward

    r = scale * (icemove_w * rets[0] + energy_w * rets[1] + defect_w * rets[2] + baseline)

• Accepted reward

    r = +1.0 * (loop_length / 4.0)

rets[0]: +1/-1 accpet/reject
rets[1]: difference of totoal mean energy 
rets[2]: difference of defect density
rets[3]: ratio of configuration difference

observations

stacked scene

• spin configuration
• trajectory
• energy map
• defect map in format of NHWC

Callable interface from libicegame

List in sqice.hpp

TODO

• Long and short loop detection mechanism
• Reward design
• Area reward
• Save loop config when accepted
• Fix ambiguous rets
• demo codes
• Analysis tools